Introduction

The human immunodeficiency virus 1 (HIV) was first introduced into the human population in the beginning of the 20th century through a zoonotic transmission event [1, 2]. Its silent spread in the following decades resulted in a globally established HIV epidemic, disproportionally affecting sub-Saharan Africa [3]. In addition to directly related healthcare challenges, the high prevalence of HIV in these countries has contributed to a strong rise in tuberculosis (TB) incidence rates [4–9]. Accordingly, HIV co-infection in Mycobacterium tuberculosis (Mtb)-infected patients has been associated with an increased risk of progression to active TB disease, an increased risk of recurrent Mtb infection, and an increased TB case-fatality rate [5, 9–13]. Despite our incomplete understanding of the interactions between TB, HIV, and the human immune system, it is widely accepted that the depletion of CD4+ T-cells underlies the high TB susceptibility and mortality in HIV-infected individuals [9, 11, 14–16].

While many studies support this increased susceptibility to TB disease, the effects of HIV co-infection on the generation of secondary TB cases remain poorly understood [9, 17]. The HIV-associated reduction in CD4+ T-cell count has been shown to be associated with an altered TB disease presentation, including lower levels of lung cavitation, lower bacterial loads in the sputum, and a higher likelihood of extrapulmonary TB [9, 18]. This distinct lung pathology could result in reduced Mtb transmission, as transmission is mainly driven by the formation of aerosols from infected lungs, and lung cavitations are known to enhance transmission [19, 20].

Several household contact studies have indicated reduced infectiousness of HIV-positive TB index cases [21–26], although it should be noted that many of these studies only considered sputum smear-negative HIV patients or patients with considerably reduced CD4+ T-cell counts [23–26]. In contrast, a meta-analysis [27] and a more recent whole-genome sequence analysis of multidrug-resistant (MDR) Mtb isolates [28] found no association between HIV co-infection and the probability of Mtb transmission, although this could result from the low sensitivity of the methodology used [29].

While the altered lung pathology of HIV co-infected TB cases could affect the rate at which these individuals transmit Mtb, the number of secondary cases generated is also determined by how long these individuals remain infectious for Mtb. Previous studies on the duration of Mtb infectiousness in HIV-positive patients showed mixed results. Several studies indicate that HIV co-infection shortens the Mtb infectious period [30–32]. This could be explained by a higher TB mortality rate or faster TB disease progression, resulting in a more timely diagnosis and initiation of treatment [8, 9, 13, 33, 34]. In contrast, one study estimated that the time until accessing TB treatment was longer for HIV-positive patients [35], which could be explained by a postponed diagnosis due to barriers to care for HIV patients, or by higher rates of smear-negativity and an atypical disease presentation.

As the effects of HIV co-infection on the Mtb transmission rate and infectious period are not well established, it remains unclear how HIV influences the overall transmissibility of Mtb. Here we analyzed four Mtb genomic datasets from sub-Saharan African countries with a high burden of HIV (Malawi, South Africa, Tanzania, and Uganda). In particular, we applied a Bayesian phylodynamic model, coupling an epidemiological model with a model of sequence evolution, to investigate how HIV co-infection affects the Mtb transmission dynamics. Our phylodynamic model stratifies the TB patient population by HIV infection status, and is parametrized with the aim of estimating the effect of HIV co-infection on both the risk of developing active TB upon exposure and the average number of secondary TB cases generated per infected individual (i.e., the effective reproductive number). Our results confirm that HIV co-infection is associated with an increased risk of developing active TB, and at the same time provide evidence for reduced Mtb transmissibility from HIV-positive TB cases.

Results

We analyzed complete Mtb genomes collected from TB patients in four sub-Saharan African countries: 1,209 sequences from Karonga District, Malawi (1995–2011) [36], 1,133 sequences from Khayelitsha, Cape Town, South Africa (2008–2018) [37], 1,074 sequences from Temeke District, Dar es Salaam, Tanzania (2013–2019) [38], and 185 sequences from Kampala, Uganda (1995–2012) [39, 40] (see Materials and methods for details on the study populations). The sequences from Uganda have been used partially in other studies [41, 42] but are analyzed here together for the first time. Mtb lineage distributions per sampling location are shown in S1 Table.

To quantify the effects of HIV co-infection on Mtb transmission in these locations, we customized a phylodynamic model based on the structured birth-death model [43], with the TB patient population stratified by HIV infection status as determined at the time of TB diagnosis (S1 Fig). In this model, transmission of Mtb within and between subpopulations (i.e., HIV-negative and HIV-positive TB cases) is described with different transmission rates (number of transmission events per infected individual per unit of time), and each subpopulation is additionally characterized by a rate of becoming uninfectious (1/infectious period) and a sampling rate. The ratio of the transmission rate and the becoming uninfectious rate corresponds to the effective reproductive number (R_e), representing the average number of secondary TB cases that one infected individual generates in the same or the other subpopulation. To explicitly model the effects of HIV, we reparametrized this model with (1) a base R_e, corresponding to the R_e for TB in a purely HIV-negative population (), (2) a parameter for the multiplicative effect of HIV co-infection on the R_e for TB, at the donor side of transmission (f₁), (3) a parameter for the multiplicative effect of HIV co-infection on the risk of getting diagnosed for active TB disease after contact with an infectious individual (f₂), (4) the rate at which HIV-negative individuals become Mtb uninfectious (δ⁻), (5) a parameter for the multiplicative effect of HIV co-infection on the rate of becoming Mtb uninfectious (f₃), and (6) a parameter for the HIV prevalence in the general population (p_HIV) (S1 Fig; see Materials and methods for more details on the phylodynamic model). The HIV prevalence was included to account for the HIV-negative and HIV-positive susceptible population sizes, which in turn influence the contact probabilities and thus the R_e in each location. HIV prevalences were not estimated from the genomic data, but set to time-varying levels based on location-specific prevalence data from World Bank [44–47] (S2 Fig). To improve the identifiability of the parameters of interest, the becoming uninfectious rate for HIV-negative individuals was fixed to 1 year⁻¹. This corresponds to an average Mtb infectious period of 1 year, which is within the range of previous estimates [30–32, 35]. We assumed no migration in the model, implying that the HIV status does not change during the course of an Mtb infection. We also assumed a constant , as justified by the results of an unstructured birth-death skyline analysis [48] (S3 Fig). Each sampling location was analyzed independently. For each location, we inferred Mtb lineage-specific phylogenetic trees, with each tree modelled as having an independent origin and evolutionary parameters. Epidemiological parameters were assumed to be the same for all lineages co-circulating within a given location. All parameters were estimated with the Bayesian phylogenetics package BEAST2 [49, 50], with prior distributions summarized in S2 Table.

The posterior maximum clade credibility trees of Mtb isolates from Tanzania are displayed in Fig 1 (trees for the other locations are shown in S4–S6 Figs), indicating limited clustering of HIV-positive patients. The posterior distributions of the parameter estimates show that the TB disease development risk for HIV-positive relative to HIV-negative individuals is significantly higher than 1 in each sampled location, with posterior means ranging from 4.48 to 11.49 (Fig 2A). These results suggest that HIV-positive individuals have a 4- to 11-fold increased risk of developing active TB upon exposure, which is in accordance with the increased TB incidence rate observed in HIV patients [9, 10, 12, 13], and can thus be considered as a ‘positive control’ in our analyses. Posterior estimates for the relative R_e (HIV-positive relative to HIV-negative individuals) are all significantly lower than 1, with posterior means ranging from 0.047 to 0.20, implying that HIV co-infection is associated with a 5- to 21-fold reduction in R_e for TB (Fig 2B). This could result from an altered TB disease presentation, as supported by the significantly lower chest X-ray scores of HIV-positive compared to HIV-negative TB patients in Uganda (Welch’s two-sample t-test, p = 0.044). Similarly, HIV-positive TB patients from Tanzania showed reduced chest X-ray scores (Welch’s two-sample t-test, p < 0.001), less cavity development (χ²-test, p < 0.001), and lower bacterial loads in the sputum (Welch’s two-sample t-test, p = 0.0027) [38], reflecting a distinct lung pathology, consistent with reduced infectiousness. Furthermore, clinical data from Malawi and South Africa showed that HIV-positive patients were strongly associated with extrapulmonary TB (χ²-test, p < 0.001 for both datasets), which is non-transmissible. All these observations are consistent with previous studies [9, 18].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Posterior maximum clade credibility trees of Mtb isolates from Tanzania.

Posterior maximum clade credibility trees per lineage, summarizing the posterior tree distribution resulting from the phylodynamic analyses on the Mtb sequences from Tanzania, with tips labeled by HIV infection status. Trees of isolates from the other locations are shown in S4–S6 Figs.

https://doi.org/10.1371/journal.ppat.1011675.g001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Phylodynamic estimates of the effects of HIV co-infection on Mtb transmission.

Prior (grey) and posterior (coloured) distributions per sampling location of the estimates for a) the relative risk of developing active TB upon exposure (HIV-positive relative to HIV-negative individuals), b) the relative R_e for TB. For all posterior distributions, the 95% HPD intervals do not contain 1.

https://doi.org/10.1371/journal.ppat.1011675.g002

We could not identify a clear effect of HIV co-infection on the Mtb infectious period, due to wide posterior distributions for some locations and conflicting results for others (S7 Fig). However, irrespective of the infectious period, HIV co-infection was associated with a reduced Mtb transmission rate (S7 Fig), suggesting that HIV mainly affects the strength, rather than the duration of infectiousness.

To investigate how much our results are impacted by the assumptions on the becoming uninfectious rate of HIV-negative individuals, we repeated the analyses with fixed values of 0.5 year⁻¹ and 2 year⁻¹ for this parameter (S8 Fig). Further sensitivity analyses comprised different priors on the parameters for the HIV effects on Mtb transmission and TB disease development risk (S9 Fig), fixing the clock rate to either 10⁻⁸ or 10⁻⁷ substitutions per site per year (S10 Fig; see S3 Table for the clock rate estimates resulting from the main analyses), setting the HIV prevalence to local instead of national estimates when available (S11 Fig; see Materials and methods), and assuming a twofold undersampling (e.g., due to decreased culturability) or oversampling (e.g., due to increased progression risk and better access to diagnosis) of HIV-positive compared to HIV-negative TB cases (S12 Fig). While the absolute values of the posterior estimates of interest were weakly dependent on the prior assumptions, all sensitivity analyses resulted in the same qualitative conclusions regarding the relative progression risk and relative R_e.

Phylodynamic birth-death estimates are not only informed by the genomic data, but also by the distribution of sample collection dates. An additional set of analyses where the sequences were ignored showed that estimates of the relative progression risk with and without the genomic data were in close agreement, with only slightly shifted posterior distributions and more certainty in the estimates when the sequences were included (S13 Fig), suggesting that they contain little information about this parameter. For the relative R_e, including the sequences resulted in shifted and narrower posterior distributions (S13 Fig), indicating that while the isolation dates and HIV infection status of patients are the major source of information, the genomic data further inform this parameter.

To identify potential biases introduced through model assumptions and priors, we repeated the analyses on datasets where the HIV status of the patients was permuted. These datasets still contained signal for the relative progression risk and relative R_e (S14 Fig). This can be explained by the HIV prevalence being 4 to 7 times higher in the sampled TB patients than in the general population in the countries under study. Indeed, randomly assigning the HIV status using the average HIV frequency in the general population during the sampling period resulted in posterior distributions for the relative progression risk and relative R_e that overlap with 1, implying no effect of HIV co-infection (S14 Fig). Together, these results demonstrate that the signal for the HIV effect parameters originates from the data rather than from model assumptions, with the parameter inference presumably being driven by 1) the high prevalence of HIV within the population of TB patients, and 2) the sampling dates and sequences informing the overall R_e estimate, which in turn constrains the HIV effect parameters.

HIV patients might show different levels of CD4+ T-cell depletion, depending on the stage of HIV infection and whether the patient is on antiretroviral therapy (ART). Several studies have indicated that decreased CD4+ T-cell counts are associated with a reduced frequency of lung cavitations (see [9] for an overview), and we found a similar association in HIV patients from Uganda (Welch’s two-sample t-test, p = 0.030), suggesting that CD4+ T-cell counts might be a better predictor of Mtb transmissibility and TB progression than the HIV infection status alone. In contrast, one study showed that TB incidence rates were increased even in HIV patients with high CD4+ T-cell counts [11], suggesting that other aspects of HIV infection might also play a role. To investigate the contribution of CD4+ T-cell counts to the observed effects of HIV co-infection on Mtb transmission, we repeated our phylodynamic analyses on sequences from South Africa and Uganda, with subpopulations defined by CD4+ T-cell counts (lower resp. higher than 350 cells/μl, the threshold recommended by WHO to prioritize patients for ART) instead of HIV status (S15 Fig). For the other sampling locations, CD4+ T-cell counts were not available. These analyses resulted in similar posterior means and HPD intervals for the parameters of interest (Fig 3), suggesting that the CD4+ T-cell count can be used as a predictor of Mtb transmission, but that, within the context of our analyses, it is not more informative than the HIV status.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. CD4+ T-cell count as predictor of Mtb transmissibility and TB disease progression.

Prior (grey) and posterior (coloured) distributions per sampling location of the estimates for a) the relative risk of developing active TB upon exposure (individuals with low CD4+ T-cell count relative to individuals with high CD4+ T-cell count), b) the relative R_e for TB. A threshold of 350 CD4+ T-cells/μl was used to classify patients. For all posterior distributions, the 95% HPD intervals do not contain 1.

https://doi.org/10.1371/journal.ppat.1011675.g003

Discussion

The effects of HIV co-infection on Mtb transmission have remained elusive, with previous studies yielding contradictory results. Here we used a phylodynamic approach to address the question through Mtb sequences and HIV/Mtb co-infection data sampled in four different African countries. Our phylodynamic analyses support previous studies showing that HIV/Mtb co-infected individuals are at high risk of developing active TB disease compared to HIV-negative Mtb-infected individuals. This observation serves as a ‘positive control’ that supports the validity our method. Moreover, we found that HIV-positive TB cases on average cause significantly fewer secondary TB cases compared to HIV-negative TB cases. These findings were reproduced independently across all four countries included in our analysis.

The finding from this and previous studies that HIV co-infection is a strong risk factor for developing active TB disease upon exposure explains why many TB epidemics in sub-Saharan Africa seem to be driven by the high HIV prevalence [4–9]. The underlying cause of this increased susceptibility to TB might be the depletion of CD4+ T-cells in HIV patients [9, 11, 14, 15]. As we could only investigate the overall risk of developing active TB disease after contact with an infectious individual, it remains unclear, based on our data, whether HIV also affects the risk of Mtb infection.

The consistently reduced TB R_e from HIV-positive individuals, observed in all countries under study, seems to be linked to a reduced number of transmission events per infected individual per unit of time. These findings are in accordance with the reduced Mtb infectiousness of HIV patients previously observed in various household contact studies [21–26]. A reduced infectiousness of HIV-positive TB cases can potentially be explained by an altered TB disease presentation in HIV-positive individuals, which could in turn result from an impaired immune system. This notion is supported by significant associations between HIV infection status and clinical variables related to lung damage and bacterial burden, observed in this and previous studies [9, 18].

The TB R_e of HIV-positive individuals might be additionally reduced through a shorter infectious period, due to more rapid disease progression and/or an increased mortality rate [8, 9, 13, 33, 34]. In contrast, increased bacterial drug resistance, delayed diagnosis due to an atypical disease presentation, and barriers to care for HIV patients could increase the Mtb infectious period of HIV-positive TB cases [35]. While we could not identify a consistent HIV effect on the infectious period, we showed that the reduced R_e of HIV-positive TB cases was linked to a lower transmission rate, irrespective of the duration of the infectious period.

The effects of HIV co-infection might be complicated by ART, which alleviates the CD4+ T-cell depletion in HIV patients [51, 52]. As information on ART was lacking for most patients, we could not directly take this into account in our analyses. However, we assumed that the CD4+ T-cell measurements from the patients in South Africa and Uganda would reflect differences in ART. As the CD4+ T-cell count classification (low/high) for these patients largely overlapped with their HIV infection status (S15 Fig), the observed effects of low CD4+ T-cell counts were largely similar to the effects of HIV co-infection. ART only recently became widely accessible to HIV patients in South Africa and Uganda [53, 54], which might explain why CD4+ T-cell count and HIV status are largely redundant. In other settings, ART might play a more important role in determining CD4+ T-cell counts, and consequently, CD4+ T-cell counts might be a better predictor of Mtb transmissibility.

HIV co-infection has been shown to be associated with rifampicin-resistant TB, potentially due to increased resistance acquisition during TB treatment [55]. While the prevalence of drug resistance during the sampling period was low in Malawi, Tanzania, and Uganda, the South-African dataset consists of rifampicin-resistant Mtb isolates only, indicating that HIV is associated with reduced Mtb transmissibility irrespective of drug resistance. Other potential confounders are poverty-related risk factors, patient sex, and patient age. However, with our current approach, these confounders are challenging to control for due to the rapidly increasing model complexity.

Notably, the major source of information on HIV effects in the model are the sampling dates and the HIV status of the TB patients, indicating that our customized birth-death model would have been able to capture most of the signal in the data even in the absence of Mtb sequences. A potential explanation for this limited signal in the sequence data is the fact that HIV status is a host-related factor that is not associated to the bacterial genetic background. Consequently, the Mtb genomes of HIV-positive TB patients are dispersed across the phylogenetic tree, resulting in high uncertainty on ancestral states and thus few informative branching events.

One limitation of our approach is that we did not account for multiple Mtb introductions into the study populations, nor for changes in the TB R_e over time. However, HIV was most likely introduced only after the establishment of different Mtb lineages, and the majority of branching events informing the epidemiological parameters occur after the introduction of HIV, suggesting minimal impact on the estimates. In support of this notion, current evidence indicates that the main Mtb lineages circulating in these parts of Africa were introduced several centuries ago [38, 41, 56–59]. Moreover, no biases were observed when removing all information about HIV status (including prevalence) through randomization of the datasets. A second limitation is that our estimates might be biased due to an underestimation of the HIV prevalence in the general population. As this potential bias originates from the input data rather than the model, it cannot be identified with our randomization approach. A third limitation is the assumption that the probabilities of contact between and within HIV-positive and HIV-negative subpopulations are solely a function of the size of each subpopulation, ignoring any preferential contacts due to social effects. Finally, a fourth limitation of the model is the assumption that individuals infected with Mtb immediately become infectious (i.e., no period of latent infection). Although this assumption could affect the interpretation of the transmission rate and infectious period, the relative R_e is expected to be robust, even if the duration of the latent period would be associated with HIV infection. Notably, our model does not distinguish between Mtb-uninfected individuals and individuals who are infected but never develop active TB disease.

Taken together, our results demonstrate that a high HIV prevalence can fuel a TB epidemic by increasing the risk for TB disease progression in HIV-positive individuals, but that these individuals do not proportionally contribute to further Mtb transmission. HIV-positive TB cases can thus be considered as ‘sinks’ in transmission chains. By contrast, HIV-negative TB cases serve as the ‘sources’, being disproportionally responsible for Mtb transmission. Our findings have implications for TB control, and call for a particular attention to HIV-negative TB cases, ideally through active case finding, thereby ensuring that these cases are diagnosed and treated as early as possible to prevent further spread of the disease.

Materials and methods

Supporting information

Acknowledgments

Calculations were performed on the Euler cluster at ETH Zürich and at sciCORE (https://scicore.unibas.ch/) scientific computing core facility at the University of Basel. We would like to thank Louis du Plessis for valuable feedback on the manuscript, Timothy G. Vaughan for help with the phylodynamic analyses, Cecilia Valenzuela Agüí and Ailene MacPherson for insightful discussions, and the Malawi Epidemiology and Intervention Research Unit for providing patient data.